Yesterday, in part one of our in-depth interview with Johns Hopkins Medicine’s Dr. Brian Hasselfeld, senior medical director of digital health and innovation and associate director of Johns Hopkins inHealth, discussed about the role of artificial intelligence in health care in general.
Today, Hasselfeld, who is also a primary care physician in internal medicine and pediatrics at Johns Hopkins Community Physicians, shifts his focus to Johns Hopkins itself, where he and several teams across the organization have deployed AI in Patient portal and ambient record applications. They are partnering with EHR giant Epic to deploy AI for chart summarization – a major step forward.
Q. Let’s move to AI at Johns Hopkins Medicine. You are using ambient recording technology. How is this working in your workflow and what kind of results are you seeing?
ONE. Definitely a very topical space. We’re seeing a number of products take on a variety of strategies. We, like many who have taken some initial steps in this space, realized that technology really wasn’t doing what it was supposed to do in healthcare.
Arguably, most data would say that, at least for clinicians, technology has been more harmful in some ways, at least to our workflow and experience in health care. So we’re trying to think of some parts where we can bring technology back to the center and make it more interesting.
Once again, many have acknowledged the burden of documentation being placed on our clinicians with the explosion of EHR content, both by regulatory requirements and common workflows across multiple systems. main system. So for most of our systems there is selected on Ambient AI, a listening device, its ambient part is listening to a clinical encounter, whether it is an outpatient visit, emergency history or inpatient stays.
And on the back end, the AI engine, often called a large language model, such as GPT, then takes the speech between multiple parties and builds it into a new paragraph.
It uses the actual functionality of those large language models to create a piece of content, usually around a specific prompt. With that model, “Let’s write a history based on this medical foundation.” And we’ve now rolled this out across a number of ambulatory or outpatient clinics, across a number of different specialty areas, now with our first product and probably thinking about how we Use multiple products to understand different levels of functionality.
I myself just went into the clinic this morning and was fortunate to be able to use the ambient AI technology with a device, my own smartphone, with our EHR on the phone and be able to launch the ambient AI product around, listen to the encounter and create a sketch, which of course, I am responsible for and need to review and edit myself to ensure clinical accuracy. It really makes clinical interactions much better.
The ability to take your hands off the keyboard, look directly at the patient and have an open conversation about a very intimate topic, their personal health and really take your eyes off the computer and back to the patient, in mind I, is the main benefit so far.
H. Johns Hopkins Medicine is also using AI to respond to message drafts on patient portals. Please explain how doctors and nurses use this method and what kind of results they achieve.
ONE. This enterprise tool is available to early adopters. Many who follow HIMSS Media content are probably aware that patient email or in-cart messages, those generated through patient portals, have exploded during the pandemic.
At Hopkins, we’ve seen the number of messages patients send to our clinicians nearly triple since pre-COVID at the end of 2019 compared to the current rate we’re seeing. And some of that is actually a good thing. We want our patients to engage with us. We want to know when they feel well or unwell and help triage.
But again, clinical workflows, including payment models and clinical care models, are not built for this constant communication, this constant communication. It is built around visits. We did something in good faith, strengthening the connection with our patients. It’s a very easy method, something we all do every day – email and texting.
We are used to communicating what we call asynchronous or through written communication. But we can’t really change the other side of it. The unintended consequence was to dump all that volume into an unchanged clinical practice system.
Now we’re all trying to figure out how to speed up improvement in the meaningful area of clinician burnout while maintaining the patient benefits of exposure more freedom with their clinical team.
So a message arrives. Some things are excluded, especially if they have attachments and the like, because those types of messages are harder to interpret. And when the message reaches a clinical care team member, those with access to the pilot implementation of AI draft responses will see the option to select a draft response based on the content of the original message, then look at the draft of the large language model response, based on some of the instructions given to try to interpret it in an appropriate way.
As a clinician, I can choose to start with that draft or start with a blank message. Stanford just put out a paper on this and it’s quite clear about some of the pros and cons, that one of the benefits is reducing the cognitive load of trying to think about responses to very common types of messages. .
We’re also seeing that clinicians who have chosen this tool and are using it on a regular basis are definitely showing a reduction in cart burnout and clinician wellness metrics. But at the same time, I think we’ve saved minimal time now because the draft responses are only really applicable and really useful to patient messaging in a small number of cases. fit. In Stanford’s published paper, that rate was 20%.
We see our clinics ranging from low single-digit percentages to 30-40%, depending on the type of user, but still less than half. The tools aren’t perfect, the workflow isn’t perfect, and it’s going to be part of that rapid iteration to figure out how we apply these tools to the most useful situations at this moment.
Q. I understand that Johns Hopkins Medicine is working on chart summarization through AI, with an initial focus on hospital residency course summaries. How will AI work here and what should you expect?
ONE. Among all the projects, this one is in its early stages. It’s a great example of the difference in technology adoption across the continuum of care and the depth of the problem being addressed.
In the previous examples, environment responses and cart drafts, we were actually working on a very concise transactional component for clinical continuity. Single visit and related discussion, unique message and response drafting. That’s very contained data.
When we start thinking about the broader topic of chart summarization, unfortunately or fortunately, the problem to be solved is limitless – the depth of the data that needs to be understood. And again, that needs to be moved from unstructured to structured.
Really, the work that we as clinicians do every time we interact with a chart, we move through the chart in a variety of ways, we extract what we feel we need to know and we summarize. It’s a complicated task. We’re trying to work in the most targeted area, during the inpatient admission, where you’re essentially more time-bound than in other versions of chart abstraction.
In outpatients, you may have to chart summarizing 10 years of information depending on why you see that clinician or why you came in. This morning I had a new patient. I need to know everything about their medical history. It’s a big graph summarization task.
In inpatient, we have the opportunity to create some time constraints around what to summarize. So, don’t even start with the whole hospitalization thing – this can actually include the reason for the hospitalization, which can then go back to the rest of the chart.
During your hospitalization, we have the daily progression of your journey during your hospital stay and changing time periods. Those issues are addressed in daily progress notes, in handouts between clinical teams. And we can narrow down the information to summarize what has changed and happened from yesterday to today, although it’s a lot of things that could happen – images, labs, notes from main team, notes from consultants, notes from nursing team.
It is more time-bound and still delivers meaningful results for inpatient groups, while also certainly identifying a well-known area of risk, which is handovers. Whenever your clinical team changes during your inpatient stay, which often happens because we do not require clinicians to work 72 hours continuously in most cases, then We have an opportunity to support those high-risk transition areas.
So when trying to limit the scope, and even here in the case of this scope limitation, there is a lot of work to be done to get a potential tool ready for practical use in Clinical workflow is, quite frankly, given the breadth and depth of available data. We’re just beginning this journey of discovery, working with our EHR partners at Epic, and look forward to seeing what’s possible here.
To watch this video interview with BONUS CONTENT not included in this story, click here.
Editor’s note: This is the seventh in a series of articles featuring leading voices in health IT discussing the use of artificial intelligence in healthcare. To read the first article about Dr. John Halamka at the Mayo Clinic, click here. To read the second interview with Dr. Aalpen Patel at Geisinger, click here. To read the third part, with Meditech’s Helen Waters, click here. To read the fourth part, with Epic’s Sumit Rana, click here. To read part five, with Dr. Rebecca G. Mishuris of Mass General Brigham, click here. And to read part six, with Dr. Melek Somai of Froedtert Health Network & Medical College of Wisconsin, click here.
Follow Bill’s HIT news on LinkedIn: Bill Siwicki
Email him: [email protected]
Healthcare IT News is a publication of HIMSS Media.
